Your question sits at the center of the AI safety debate — I'll compare the risks and benefits of open versus closed models.
Neither is inherently safer — they are safer and riskier in different ways, and the evidence on net risk is still mixed.
The short answer
> The evidence is genuinely mixed. Open weights raise proliferation concerns, because they can be copied widely and have their safety training removed. But openness also enables independent auditing, red-teaming, and interpretability research that closed APIs block
Studies that tried to quantify the difference have not found a clear winner.
•The 2024 US National Telecommunications and Information Administration (NTIA) report framed this as a question of marginal risk - the risk unique to open-weight models relative to closed models and existing technology - and found the evidence too immature to conclude that open models pose more marginal risk than benefit. It recommended monitoring rather than restriction.
•A Stanford Institute for Human-Centered AI study reached a similar conclusion: there is currently limited evidence of the marginal risk of open foundation models compared to closed models or existing technologies, while noting open models provide significant benefits by combatting market concentration, catalyzing innovation, and improving transparency.
Why open-source / open-weight can be safer
1.Transparency and independent oversight: Open-weight systems increase transparency, allow for widespread red teaming, and decrease market concentration. The case for openness is that the same transparency enables safety: independent researchers can audit, red-team, find vulnerabilities, and study interpretability in ways a closed API forbids.
2.Faster vulnerability patching: Empirical data from software development proves that open-source systems identify and patch vulnerabilities significantly faster than siloed, proprietary alternatives. When we decentralize AI, we distribute oversight across a global community rather than leaving safety in the hands of a few executives.
3.Avoiding single points of failure and power concentration: From this perspective, open-source AI is actually safer because it eliminates the single point of failure and prevents a small group of engineers from having unilateral control over powerful cognitive tools.
4.Eliminating black-box risks: Closed training data creates a "black box" of potential problems, and security professionals warn against security-by-obscurity. When critical parts of development are kept secret, we simply cannot be sure of its safety.
Why open-source / open-weight can be riskier
1.Irreversible proliferation: Open-weight models can allow harmful AI capabilities to proliferate rapidly and irreversibly. Just as they can be modified for beneficial uses, they can also be modified for harmful ones.
2.Guardrails can be removed quickly and cheaply: Safeguards designed to prevent harmful behaviours can also be quickly and cheaply removed from open-weight models. Current fine-tuning techniques can be undone using just dozens of training examples in minutes, and tamper-resistant fine-tuning is still a frontier problem.
3.Lower barrier for misuse: The NTIA noted concern that open weights could make it easier for non-experts to misuse models, including toward dangerous ends. Jailbreaks and guardrail bypasses are easier on open models - when someone can spin up an uncensored chatbot with zero oversight, risks escalate fast.
4.Perceived risk: In surveys, respondents consider open source AI tools to be riskier than proprietary AI for most types of AI risks, with the most cited risks being cybersecurity (62%), regulatory compliance (54%), and intellectual property (50%).
Closed models are not immune - there is evidence you can take models, closed or open, and hack them to remove their guardrails - but with closed models the provider retains server-side controls.
Why closed/proprietary can be safer and riskier
Safer aspects of closed:
•The deployer can draw from a rich toolkit for making it safer, such as content filters, and fully controls the points of access, allowing enforcement of acceptable usage policies.
•Safety protections operate server-side and cannot be modified by end users, unlike open-weight models where weights are accessible.
•Once a system is released with open weights, it cannot be rolled back, while halting download access can still slow proliferation for closed systems.
Riskier aspects of closed:
•Opacity prevents external verification, creates vendor lock-in, and concentrates power, which critics argue is riskier to rely on.
•Lack of self-hosting means prompts and data must leave your infrastructure, raising privacy and data governance risks.
Practical distinction
Most models marketed as "open source" today are actually open-weight (weights downloadable but training data/recipe private) rather than fully open-source (weights + data + code). That middle tier matters for what you can audit, fine-tune, and self-host versus what you rent as an API.
Bottom line:
•If your threat model is misuse by bad actors after release, closed is easier to contain in the near term.
•If your threat model is undetected flaws, bias, backdoors, concentration of power, and long-term resilience, open is safer because it can be inspected and fixed by anyone.
Most governments and researchers currently treat this as an unresolved marginal-risk question to be managed with layered mitigations - data curation, tamper-resistant training, audits, monitoring tools, and staged deployment - rather than a binary safer/riskier label.